When memory is sold as a tier

The paid plan promises better memory, and you cannot tell what you would be buying. Memory is the one feature in these products where the marketing word and the engineering reality diverge most, and it is also the feature most often placed behind the highest tier. Both of those facts have the same cause.

Memory is the feature with a recurring bill

Most features cost money once, to build. Memory costs money on every single turn, forever, for every user who has it.

The reason is that whatever the app remembers has to be sent to the model again with each new message. A longer memory means more text in every request, and the model is billed by the amount of text. Illustratively: if a bare conversation costs a fifth of a cent per reply and a generous memory doubles the text going in, the cost per reply roughly doubles too — every reply, for as long as the account exists. Those numbers are round and hypothetical, but the structure is the point. Memory is not a feature with a price; it is a multiplier on the cost of everything else.

That is why it sits in a tier. An operator cannot give unlimited memory to free accounts without committing to an open-ended monthly expense per free user, and the accountancy of that decides the pricing page.

What a memory tier actually buys

There are only a few ways to build this, and each one loses something different. A tier upgrade nearly always means more of one of them rather than a new kind of remembering.

A longer recent window — more of the actual conversation is sent verbatim. Accurate for what it covers, and it still ends abruptly at a boundary you cannot see.

A longer or more frequent summary — more history is compressed into a paragraph. Cheap, and it keeps the gist while dropping specifics, which is why a summarised week feels remembered but the details inside it are gone.

More extracted facts — a bigger list of stored details. Precise about what is on the list and blind to what was never added to it.

Search over past messages — old text is retrieved when it looks relevant. Excellent when the retrieval finds the right passage, and silently absent when it does not.

The upgrade therefore does not remove the loss. It moves the boundary. A tier that doubles the window gives you twice as much verbatim history and one boundary instead of one boundary; a tier that adds fact extraction gives you reliable recall of listed items and nothing else. Which one you are buying is a concrete question, and it is usually not answered on the pricing page.

The purchase you cannot test first

This is the awkward part. Memory quality is only observable over weeks, because the failure mode is a detail dropping out of a conversation you had a fortnight ago. A free trial short enough to be cheap for the operator is also too short to demonstrate the thing being sold.

So the tier is bought on a description, and the descriptions are adjectives: enhanced, long-term, persistent, unlimited. “Remembers everything” is a claim about a feature, and the useful version of the question is always what shape the remembering has — how much verbatim, how compressed, retrieved how, and what happens at the edge.

THE PRODUCT — memory as a paid tier

  · "Enhanced memory" on the paid plan
                    → usually more of one existing mechanism:
                      window, summary, fact list, or search.

  · Why it is priced at all
                    → it is re-paid on every turn. Memory is
                      a multiplier on per-reply cost, not a
                      one-off feature.

  · "Remembers everything"
                    → an adjective. The question is the
                      shape: how much verbatim, compressed
                      how, retrieved how, ending where.

  · How long memory is, per tier
                    → THE OPERATOR DECIDES, and can shorten
                      it for cost reasons without an
                      announcement.

  · What a downgrade does to stored memory
                    → CHECK THE POLICY and the plan page.
                      Truncated, hidden, or deleted are three
                      different answers.

What you can check

Ask the pricing page for a unit. A tier that states a number — messages of history, a count of stored facts, a time span — is describing a product. A tier that states an adjective is describing a feeling. The first can be compared between plans; the second cannot.

Look for whether stored memory survives a downgrade. Memory built up on a paid plan may be truncated to the free length, kept but inaccessible, or deleted. All three are reasonable engineering choices and they are very different outcomes for you. The plan page or the help documentation usually says, and if neither does, that is the finding.

Test the boundary deliberately, early. Mention something specific and unusual, then ask about it a week later. That tells you more about the shape of the memory than any tier description, and it costs nothing.

Separate memory from continuity of character. A character can hold its manner perfectly while having forgotten your facts, and it can recite your facts while behaving like a different character. They are separate systems and separate purchases.

What this doesn’t tell you

It does not tell you what any app’s memory tier contains. That is per-app, undocumented in detail, and subject to revision.

It does not tell you which mechanism yours uses, because apps do not generally publish that, and the observable behaviour only narrows it down.

And it does not tell you whether the upgrade is worth the money — only that the question has a concrete form, which is better than the adjective you were offered.